Mihir Pradeep Deshmukh

I am a student at Worcester Polytechnic Institute(WPI) pursuing Robotics Engineering. I am interested in the domain of Robotics, specifically Perception, Computer Vision and deep learning in the field of manipulation and mobile robots. I am currently working on vision transformers and grasp generation.

I am currently seeking full time opportunities from Summer’25

Academic Background

• Aug 2023 - May. 2025: M.S. in Robotics Engineering, Worcester Polytechnic Institute (WPI)
• Aug 2018 - May. 2022: BTech in Electronics and Telecommunication Engineering
  Minor in Computer Engineering,
  College of Engineering, Pune (COEP)

Research Interests

• Computer Vision: Image segmentation, monocular depth estimation, object detection
• Deep Learning: CNNs, Transformers, transfer learning
• Robotics: Manipulation, grasp generation and perception

Experience

Magna Electronics - Software Algorithm Intern

• Performed Lidar–Camera extrinsic calibration to project Lidar-generated 3D bounding boxes into BEV space, enabling ground truth comparison for monocular camera-based object detection.
• Processed CAN bus data to extract GNSS information for HD map queries, enabling real-time location spoofing, lane profile reconstruction, and hardware-independent debugging environment.
• Benchmarked Autoware's planner in Carla simulation and contributed to the system design of landmark-based localization.

Bajaj Finserv - Software Engineer

• Leveraged .Net Core for Web API creation and Ado.Net for database integration as a member of the Collections Portal Team.
• Developed front-end applications in Angular.

Mitacs Globalink Research Internship (McMaster University) - Research Intern

• Worked under Dr. Gary Bone on the project ”Collaborative Robot Arm Software Development”.
• Applied concepts of Ransac segmentation, clustering, and Transforms in Webots and MATLAB. Simulated PR2 for collaboratively picking objects by processing the point cloud.

Binary Robotics - Project Intern

• Calibrated cameras using OpenCV to enhance the auto-pick and place mechanism of a gantry system.
• Designed an electronic architecture for a 5 axis Gantry system equipped with hybrid servos.

Robotics and Automation Lab, COEP - Undergraduate Researcher

• Handled the programming, circuit & PCB designing, perception, and sensor fusion for Omni-wheeled robots as well as research projects.
• Participated in ABU Robocon 2020 and 2021 and worked from ideation to prototyping the robots.

Projects

Dexterous Picking Manipulation.

• Directed Research, Advisor - Dr. Berk Calli
• Working on identifying grasp candidates for diverse object interactions like sliding and pushing in vertical and horizontal orientations.
• Generated simulation data and subsequently concentrated on developing and training models using this data, focusing on transitioning them into real-world applications.

Structure from motion(SfM) & NeRF

• Applied the chirality condition to ensure the geometric validity of triangulated points for accurate 3D reconstruction.
• Enhanced scene depth and detail by integrating additional views through Perspective-n-Points (PnP), followed by Bundle Adjustment for optimal alignment and precision of 3D points.
• Implemented Neural Radiance Fields (NeRF) to synthesize novel views of scenes from a sparse set of input images, leveraging differentiable volume rendering for photorealistic image generation.

Deep Learning based Robotic Grasping of unknown objects.

• Developed a pipeline to optimally Grasp objects of variable shape, size, and orientation using vision capabilities.
• Applied VGG16 and ResNet50 architectures through Transfer Learning in PyTorch. Adapted this to a custom 3D-printed 5-DoF robotic arm with MoveIt and the KinectV2 depth camera.

Visual Slam and Object Recognition using Kinect v2 & ROS

• Implemented RTAB map in gazebo simulator and tested the same using Kinect V2.
• Employed the YOLO v3 framework for real-time object detection in tandem with map generation.

AutoCalib

• Implemented Zhang's camera calibration method, utilizing non-linear optimization for estimating intrinsic \& distortion parameters.
• Implemented distortion model including radial distortion, with parameters estimated through iterative optimization.

Metalimbs

• Simulated an anthropomorphic robotic arm with six DOF using human foot movements, integrating leg position tracking via an IMU and flex sensor-actuated gripping.
• Utilized MoveIt for inverse kinematics and trajectory planning.

Rubik’s Cube Solver.

• Designed an automated system to solve a Rubik’s Cube, utilizing a camera for input and six stepper motors to actuate each face, powered by an STM32 microcontroller.
• Developed and optimized a color detection and cube-solving algorithm using OpenCV, with a focus on contour analysis.

Publications